[test] Expand the tolerance for testing on metal backend

[Rullec]

Related issue = Close #1679

As @archibate adviced, I expand the approx value on metal platform.

Now test_ad_basic.py can work properly on my machine!

[Click here for the format server]

---

[codecov]

Codecov Report

Merging #1779 into master will decrease coverage by 0.93%.
The diff coverage is 0.00%.

@@            Coverage Diff             @@
##           master    #1779      +/-   ##
==========================================
- Coverage   43.59%   42.66%   -0.94%     
==========================================
  Files          44       44              
  Lines        5968     6174     +206     
  Branches     1074     1075       +1     
==========================================
+ Hits         2602     2634      +32     
- Misses       3212     3384     +172     
- Partials      154      156       +2     

Impacted Files	Coverage Δ
python/taichi/testing.py	75.75% <0.00%> (-1.67%)	⬇️
python/taichi/lang/ops.py	43.89% <0.00%> (-9.09%)	⬇️
python/taichi/lang/expr.py	61.95% <0.00%> (-5.50%)	⬇️
python/taichi/lang/matrix.py	64.73% <0.00%> (-4.59%)	⬇️
python/taichi/lang/snode.py	65.83% <0.00%> (-4.44%)	⬇️
python/taichi/lang/impl.py	64.50% <0.00%> (-2.07%)	⬇️
python/taichi/lang/transformer.py	80.92% <0.00%> (-0.69%)	⬇️
python/taichi/main.py	23.57% <0.00%> (-0.39%)	⬇️
python/taichi/lang/util.py	31.36% <0.00%> (-0.38%)	⬇️
python/taichi/lang/kernel.py	70.96% <0.00%> (-0.14%)	⬇️
... and 10 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 5c39e84...ee8f1b1. Read the comment docs.

[k-ye]

Thanks this is great! Sorry it just occurred to me, but could you try running with fast_math=False (without changes in this PR)?

[Rullec]

Thanks this is great! Sorry it just occurred to me, but could you try running with fast_math=False (without changes in this PR)?

Thanks for the reply!
Just now I tried the fast_math=False option by

@if_has_autograd
@ti.all_archs_with(fast_math=False)
def test_trigonometric():
    # grad_test(lambda x: ti.tanh(x), lambda x: np.tanh(x))
    grad_test(lambda x: ti.sin(x), lambda x: np.sin(x))
    # grad_test(lambda x: ti.cos(x), lambda x: np.cos(x))
    # grad_test(lambda x: ti.acos(x), lambda x: np.arccos(x))
    # grad_test(lambda x: ti.asin(x), lambda x: np.arcsin(x))

And this test can still be wrong if I didn't expand the approx in testing.py. Log shown below:

>       assert y[0] == approx(npfunc(v))
E       assert 0.23188334703445435 == 0.2318703549116868 ± 2.3e-07
E        +  where 0.2318703549116868 ± 2.3e-07 = approx(0.2318703549116868)
E        +    where 0.2318703549116868 = <function test_trigonometric.<locals>.<lambda> at 0x111369598>(0.234)

test_ad_basics.py:47: AssertionError

Can I do more tests for you? Maybe this problem shouldn't be ignored in this way easily ;)

[k-ye]

I see.

Can I do more tests for you?

No that's the only thing i could think of... Thanks!

Maybe this problem shouldn't be ignored in this way easily ;)

Yeah, but given it didn't fail on my side, I'm attributing this to hardware differences.. ;-(